Erfana Skills
qodeca/erfana-skills
Design + orquestração de agentes em um plugin Claude Code open-source
Ver no GitHubPlugin Claude Code de código aberto para design e orquestração de agentes. Oferece 6 skills de design (protótipos, decks, animações, infográficos, críticas) e 6 skills de orquestração (gerenciamento de agentes, issues, relatórios, especificações). Tudo integrado em um único toolkit com 87 agentes compartilhados.
Instalação
npx convez-ai qodeca/erfana-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Criar protótipos UI clicáveis (iOS, Android, web, desktop)
- Gerar decks de apresentação 1920×1080 com exportação PDF/PPTX
- Produzir animações e motion graphics em MP4/GIF com áudio
- Desenhar infográficos verticais de dados em qualidade print
- Gerenciar ciclo de vida de agentes Claude e orquestrar tarefas complexas
- Revisar trabalhos de design com feedback estruturado em 5 dimensões
Veredito

Ferramenta robusta e bem documentada para quem precisa criar artefatos visuais profissionais e orquestrar múltiplos agentes Claude em fluxos complexos. GPL-3.0 permite uso em projetos open-source, mas restringe comercialização sem compartilhamento de código.
README
A tradução para o português deste README ainda está na fila.
Erfana Skills
An open-source Claude Code toolkit (GPL-3.0-only) – design and orchestration in one plugin.
by Qodeca
Two domains in one plugin:
- Design. Build production-grade design artifacts directly inside Claude Code: clickable UI prototypes, 1920×1080 slide decks (HTML / PDF / editable PPTX), MP4 / GIF motion graphics, vertical print-grade infographics, and 5-dimension design critiques.
- Orchestration. Manage Claude Code agents and skills, GitHub issues, consulting reports, articles, and 4-tier specifications. 6 orchestration skills plus 87 shared agents, all delegating substantive work via the
Tasktool.
Free and open source under the GNU General Public License v3.0 only. Contributions are welcome — see CONTRIBUTING.md, CODE_OF_CONDUCT.md, and SECURITY.md. "erfana" and "Qodeca" names and logos are trademarks; the license does not grant rights to them — see TRADEMARKS.md.
What's in this plugin
Design skills
| Skill | What it does | Triggers |
|---|---|---|
erfana:design-direction | The brief is vague – recommends 3 differentiated philosophies from a 20-school library, shows prebuilt samples, generates 3 visual demos to compare. Run this first when no visual direction is set. | "what style should I use", "recommend a style", "I don't know what style", "design advisor", "make it look good" |
erfana:design-prototype | Hi-fi clickable UI prototypes – iOS, Android, web, desktop. Single-file HTML, real device frames, real images, Playwright-verified. | "build a prototype", "iOS prototype", "Android prototype", "app mockup", "clickable design", "hi-fi mockup" |
erfana:design-slides | 1920×1080 HTML presentation decks with PDF + editable PPTX export. Multi-file aggregator for ≥10 pages, single-file deck for ≤10 pages. | "design a deck", "design a slide deck", "pitch deck", "keynote", "PPT", "editable PPTX" |
erfana:design-motion | Timeline-driven animations exported as MP4 (25fps base, 60fps interpolated) or palette-optimized GIF. Optional BGM + SFX with frequency separation. Watermark: Created with erfana. | "animate this", "motion design", "export MP4", "export GIF", "60fps video", "motion graphics" |
erfana:design-infographic | Vertical (1080×1920) print-grade data visualizations. Honest placeholders over fabricated numbers. | "infographic", "data visualization", "data viz", "vertical infographic", "chart design" |
erfana:design-review | 5-dimension critique of completed design work – Keep / Fix / Quick Wins. Severity-tagged, implementable feedback. | "design review", "critique", "rate this design", "score this", "expert review" |
Orchestration skills
| Skill | What it does | Triggers |
|---|---|---|
erfana:managing-agents | Lifecycle management for Claude Code agents – research, design, validation phases. Ships its own validation/review checklists. | "create agent", "review agent", "modify agent" |
erfana:managing-articles | End-to-end medium-form article authoring (research → outline → draft → review → publish), bilingual Polish/English. Delegates to 5 plugin-root article-* agents. | "write article", "research article", "publish article" |
erfana:managing-issues | Full GitHub-issue lifecycle: create issues, implement them through phased quality gates, and review code or PRs. Runs on main-default repos, stack-agnostic, with a skill-wide untrusted-data boundary and confirm-before-destructive git ops. See docs/architecture.md. | "create issue", "implement issue", "review code", "review PR" |
erfana:managing-reports | Professional consulting reports with Pyramid Principle, SCQA, Five Cs framework, sentence-case validation. 11 internal validation agents. | "create report", "review report", "validate report" |
erfana:managing-skills | Lifecycle management for Claude Code skills following Anthropic best practices. Includes the Modernize operation (v4.2.0+) that applies Opus 4.7 patterns to existing skills. | "create skill", "review skill", "modify skill", "modernize skill", "apply 4.7 patterns" |
erfana:managing-specs | 4-tier specification management: T1 issue, T2 spec, T3 lite, T4 standard. Delegates to plugin-root spec-* agents, with trust-boundary controls and a transactional single-writer registry (schema v3, auto-migrated on first touch). See docs/architecture.md. | "create spec", "validate spec", "T3 lite spec", "T4 standard" |
Process skills
| Skill | What it does | Triggers |
|---|---|---|
erfana:grill-me | One-at-a-time Socratic interrogation of a plan or design – walks the decision tree, recommends an answer per branch, explores the codebase first when the answer is already there. Imported from upstream superpowers:grill-me in v4.2.3. | "grill me", "stress-test this plan", "interview me about this", "challenge my plan", "poke holes in this", "walk the decision tree" |
Verification skills
| Skill | What it does | Triggers |
|---|---|---|
erfana:fact-checking | Validates markdown analysis documents against source materials (interviews, vendor docs, knowledge-base) to catch AI hallucinations – extracts atomic factual claims, traces each to its source passage, classifies findings by severity (Critical / Error / Warning / Info), and applies user-approved corrections. Five-phase orchestrator (Setup → Extraction → Verification → Interactive review → Fix application) backed by four fc-* plugin-root agents. Manual-only via /erfana:fact-checking <target-file>; not auto-discovered. Treats every ingested document as untrusted data, reconciles parallel verification by claim id, and anchors fixes on verbatim text. Migrated from a Qodeca consulting project, Modernize-passed in v4.2.7, lens-review-hardened in v4.6.0. | "fact-check this document", "verify against sources", "validate analysis", "check for hallucinations", "verify document" |
Bootstrap and shared agents
| Component | What it does |
|---|---|
using-erfana | Bootstrap. Lists available skills, establishes the 1% rule, dispatches design-vs-orchestration. Loads automatically. |
agents/ (87 shared agents) | Shared agent pool the orchestration skills delegate to via the Task tool. Prefix breakdown: spec- (23), mi- (13), ms- (10), ma- (7), article- (5), e2e- (4), fc- (4), release- (2), UI/UX (4), tech-domain (6), generic (9). |
Safety hooks (v4.1+)
Four hooks run silently in the background once the plugin is enabled, providing a project-agnostic safety net:
| Hook | When | What it does |
|---|---|---|
bash-safety | Before any Bash tool call | Blocks destructive shell patterns – rm -rf of system/home dirs, force-push to protected branches, IMDS metadata exfiltration, privilege escalation (sudo/doas), tar --absolute-names, `curl |
secret-detector | Before Write/Edit/MultiEdit | Blocks ~20 secret/token patterns before they hit disk – AWS, OpenAI, Anthropic, GitHub, GitLab, Hugging Face, Sentry, Postman, Slack, npm, Stripe, Google, Azure, database URIs, JWTs, PEM keys. Skips test fixtures, examples, markdown docs. |
post-compact-reminder | After context compaction | Re-injects load-bearing facts (temporal awareness, honesty discipline, verification rules, agent delegation) plus the current git branch + status snapshot. |
verify-completion | When the agent considers stopping | Asks the agent to keep working when it claims success without citing executed tests, exit codes, gate output, or screenshots. v4.2.9+ allowlist: messages carrying the <!-- erfana:status-template --> sentinel emitted by the status commands bypass the check; Gate 16 enforces sentinel symmetry across the two command files and the hook. Falls back to the unstripped body when the reply has an odd number of code fences so success claims after an unclosed fence stay visible. |
All four are project-agnostic – no personal style preferences. They activate only after the next Claude Code session restart following plugin install or update.
Cross-platform (v4.2.20+). Each hook ships a .sh (macOS/Linux) and a .ps1 (Windows) sibling, run through the dispatch.sh launcher so the safety net works on native Windows too (where Git Bash ships without jq). The mechanism and the one uncovered case (a Windows host with no Git Bash) are documented in docs/architecture.md.
Slash commands
| Command | What it does |
|---|---|
/erfana:doc-update | Refreshes project documentation against the current state of the code – detects the change set from the working tree, sweeps the whole doc surface (docs/, README, CHANGELOG, CLAUDE.md / AGENTS.md, API specs, ADRs), and evicts status/changelog content into its home docs. Takes no git action by default; deletions and new files are confirmed via AskUserQuestion. Flags: path-or-glob, --dry-run, --offline, --commit / --push. Full contract: commands/doc-update.md. |
/erfana:project-status | One-shot Pyramid-Principle project-status brief for a Product Owner / PM / BA audience – three axes (what we worked on, what we accomplished, where we landed) plus a recommended next step, grounded in git and GitHub state. Read-only, no side effects. Useful when context-switching across many Claude Code tabs. Full contract: commands/project-status.md. |
/erfana:session-status | The same Pyramid-Principle brief scoped to the current Claude Code session – same PO/PM/BA audience and three axes, sourced from the in-context conversation with a light git probe. Read-only; use it after a context compaction or when returning to a long-running tab. Full contract: commands/session-status.md. |
/erfana:explain-issue | Translates a single GitHub issue into a Product Owner / PM / BA brief (accepts a bare number, #N, or full URL). Pulls the issue, comments, linked PRs, and referenced files/specs to ground the translation, but stays purely descriptive – no suggested next step. Non-interactive, read-only. Full contract: commands/explain-issue.md. |
/erfana:lens-review | Researched multi-lens code review over any target – /erfana:lens-review <path | #PR | "description"> [--lens a,b,c] [--out file.md]. Fans out reviewers (up to 10, chosen at runtime), each grounded in cited best practices from the last ~12 months, then returns one severity-ranked, plain-language report (PM/PO-facing, with full technical detail kept for engineers). Manual trigger only; distinct from /review by its live research and any-target scope. Full contract: commands/lens-review.md. |
Skills auto-discovered from skills/*/SKILL.md; agents from agents/*.md (plugin root) and skills/<skill>/agents/*.md (skill-internal); hooks from hooks/hooks.json; commands from commands/*.md. Shared design content (workflow templates, content guidelines, asset bundle: 4 references / 37 SFX / 6 BGM / 24 showcases) lives in skills/design-shared/.
Heads-up: generic-name agents. The plugin ships ~15 agents with generic names – any agent in
agents/whose name does not start with a team prefix (mi-,ma-,ms-,spec-,e2e-,release-,nest-,react-). At the v4.0.0 release these includearchitecture-reviewer,bug-investigator,code-reviewer,commit-writer,refactor-advisor,security-auditor,software-developer,solution-architect,solution-reviewer,technical-architect,test-writer,ui-designer,ui-reviewer,ux-designer,ux-reviewer. They may collide with built-in Claude Code agents or with agents shipped by other plugins (superpowers:*,feature-dev:*, etc.). Last-loaded wins; behavior in mixed-plugin environments is non-deterministic. To target this plugin's copies specifically, prefer prefix-named agents in theTasktool (e.g.mi-codebase-explorer,ma-designer). Full security implications including the highest-trust shadow target are documented inSECURITY.md.
Confidentiality
Prompts you send through any skill or agent in this plugin are forwarded to Anthropic's Claude API for processing. Anthropic acts as a sub-processor of any data contained in your prompts.
Treat the plugin like any other AI tool routed through external infrastructure:
- Do not paste: client-confidential data, customer PII, financial figures marked sensitive, unreleased product specs, internal credentials, or anything covered by an NDA you signed.
- Default to public-grade brand context when describing the design target (paste your published color palette, not a draft you have not shipped).
- For sensitive design work, work locally and skip this plugin.
Review Anthropic's data-usage and privacy terms before routing sensitive material through any Claude-based tool, and treat the conservative defaults above as policy.
Install
Installing from a public marketplace needs no GitHub token or special access.
1. Add the marketplace and install
Inside Claude Code:
/plugin marketplace add qodeca/erfana-skills
/plugin install erfana@erfana-skills
The marketplace registers under the name erfana-skills (matches marketplace.json). The install line uses erfana@erfana-skills – the plugin name @ the marketplace name.
2. Verify
/plugin list
You should see erfana@erfana-skills enabled. Then type a trigger phrase like:
design a slide deck
erfana:design-slides should activate. (Or erfana:design-direction first if you want it to suggest a visual style – try "what style should I use for a Q2 retrospective deck" for that path.)
Invocation forms – bare vs. namespaced
Both forms resolve to the same skill:
/erfana:design-prototype # canonical, namespaced (use this)
/design-prototype # bare form, also works
Use the namespaced form in shared docs, screenshots, and team conversations. The bare form is a built-in Claude Code convenience – it exists for autocomplete brevity but collides with personal or other-plugin skills of the same name. The /erfana: prefix prevents that collision and makes the source explicit.
Tracked upstream: anthropics/claude-code#43695 requests a require-namespace: true flag to remove the bare form. When that ships, this plugin will adopt it.
How updates work
Third-party marketplaces like this one have auto-update off by default – only Anthropic's own marketplaces auto-update without asking (docs). So by default you pull new releases manually:
/plugin marketplace update erfana-skills
/plugin update erfana@erfana-skills
(After /plugin update, restart Claude Code to apply.)
To have releases picked up automatically at session start instead, opt in once: run /plugin, open the Marketplaces tab, select erfana-skills, and choose Enable auto-update. Org admins can set "autoUpdate": true on the marketplace's extraKnownMarketplaces entry in managed settings to enable it for everyone.
If still stale, clear the cache and restart Claude Code:
rm -rf ~/.claude/plugins/cache/
SSH-based marketplace URLs are not recommended (known Windows issue).
Pin to a specific version (optional)
If you turned on auto-update and need stability for a critical project, pin to a specific version:
/plugin install erfana@erfana-skills@v6.0.0 # for example – pick the version you want
Replace v6.0.0 with whichever release you want to lock to. A pinned version is never auto-updated. To upgrade later, run the same command with a newer tag and restart Claude Code.
Use case: you are mid-flight on a deck, a new version drops, and you do not want trigger-phrase behavior to shift under you. Pin until you are done, then unpin (/plugin install erfana@erfana-skills without a @vX.Y.Z suffix).
Typical workflows
Skills compose. The most common chains:
Pitch deck for an investor or client (visual style not yet decided)
erfana:design-direction– pick a visual philosophy from a 20-school library; see 3 demos comparederfana:design-slides– build the deck in the chosen directionerfana:design-review– score the result before sending
Sample opener: "What style should I use for a Series B pitch deck? Help me decide, then build it."
Hi-fi app prototype (iOS / Android / web)
erfana:design-prototype– build clickable single-file HTML mockup with real device frameerfana:design-review– catch hierarchy / craft / functionality issues before delivery
Sample opener: "Build me a clickable iOS prototype for a Pomodoro app, 4 screens."
Product-launch animation with audio
erfana:design-direction(only if visual style is unclear)erfana:design-motion– timeline-driven animation with BGM + SFX, MP4 / GIF exporterfana:design-review– pre-publish QA
Sample opener: "Make me a 15-second product launch animation in the Field.io kinetic style."
Vertical infographic from real data
erfana:design-direction(if no chosen philosophy)erfana:design-infographic– print-grade vertical layout with honest placeholderserfana:design-review– typography / hierarchy pass
Sample opener: "Design an infographic comparing our Q2 vs Q3 numbers."
Stress-test a plan before implementing (process skill, v4.2.3+)
erfana:grill-me– Socratic walk through the decision tree; one question at a time, recommended answer per branch, explores the codebase before asking when the answer is already encoded there- Downstream skill of choice –
erfana:managing-issuescreate /erfana:managing-specs(T1–T4) / anyerfana:design-*skill, depending on what the locked plan is meant to produce
Sample opener: "Grill me on this rollout plan before we build it – I want to ship X by Friday, here's my draft approach."
Why workflows matter: each output skill works alone, but design-direction first prevents generic AI-look outputs and design-review last catches issues humans miss. Skip neither for important deliverables.
Troubleshooting
/plugin marketplace add fails or hangs
Check network access to api.github.com and raw.githubusercontent.com (a corporate proxy may block them — see the proxy note below). The public marketplace needs no token.
Auto-updates haven't appeared in 2+ sessions
Force a refresh: /plugin marketplace update erfana-skills, then /plugin update erfana@erfana-skills, then restart Claude Code. If still stale, clear the cache: rm -rf ~/.claude/plugins/cache/ and restart.
/plugin install reports "marketplace not found"
The registered marketplace name comes from marketplace.json (erfana-skills). Make sure step 3 ran successfully – /plugin marketplace list should show it.
Trigger phrases don't activate the right sub-skill
Confirm the plugin is enabled (/plugin list) and try a trigger phrase from the skill table above. Each sub-skill (design-direction, design-prototype, design-slides, design-motion, design-infographic, design-review) has its own keyword set – if your phrasing is far from the listed triggers, mention the deliverable type explicitly ("prototype", "slide deck", "animation", "infographic").
/erfana:design-prototype and /design-prototype both appear in autocomplete
Expected behavior. Plugin skills are registered at both the bare path (/design-prototype) and the namespaced path (/erfana:design-prototype) – see Invocation forms. Use the namespaced form to avoid collisions with personal or other-plugin skills.
Plugin updated but Claude Code still uses old behavior
Force a refresh: /plugin marketplace update erfana-skills, then /plugin update erfana@erfana-skills, then restart Claude Code. If still stuck, clear the cache: rm -rf ~/.claude/plugins/cache/ and restart Claude Code.
Marketplace add hangs / silently fails behind a corporate proxy
If your network blocks api.github.com or raw.githubusercontent.com, marketplace operations fail without a clear error. Ask IT to allowlist both. As a workaround, git clone the repo locally and register it as a local-path marketplace: /plugin marketplace add /absolute/path/to/erfana-skills. Pull updates manually with git pull + /plugin marketplace update erfana-skills.
Skill scope precedence
If you have a personal or project skill that shares a name with one in this plugin (e.g. you also have ~/.claude/skills/design-slides/SKILL.md), Claude Code resolves in order: project (.claude/skills/) → personal (~/.claude/skills/) → plugin (erfana@erfana-skills). To force the plugin version, rename your personal copy or add it to a different namespace.
Local development (testing changes before they hit main) Clone the repo locally, then register it as a local-path marketplace:
/plugin marketplace add /absolute/path/to/your/clone
/plugin install erfana@erfana-skills
Iterate on your fork. When ready, push a feature/... branch (cut from develop) and open a PR against develop. Once it ships in a release, switch back to the published marketplace: /plugin marketplace remove erfana-skills then /plugin marketplace add qodeca/erfana-skills and /plugin update erfana@erfana-skills.
License
Copyright © 2025-2026 Qodeca sp. z o.o.
Licensed under the GNU General Public License v3.0 only (GPL-3.0-only). You are free to use, study, share, and modify it; distributed derivatives must remain under the same license and ship their source. The license covers code, documentation, and bundled assets — it does not grant rights to the "erfana" or "Qodeca" names or logos (see TRADEMARKS.md). Per-file licensing follows the REUSE specification; run reuse lint to verify.
Maintainer info
Repository: github.com/qodeca/erfana-skills
Maintainer: Marcin Obel (@marcinobel). Governance: GOVERNANCE.md. Support: SUPPORT.md.
Note: the
v1.0.0git tag exists but is not installable (it predates amarketplace.jsonschema fix shipped inv1.0.1). If you pin to an exact version, start fromv1.0.1or later.